If an AI companion repeatedly calls a fictional character by the wrong nickname, regenerating the latest sentence may make the screen look right without correcting the context behind the next reply. The useful Nomi–Kindroid comparison is therefore about where a correction belongs, rather than a claim that one service never forgets.
This article uses public documentation reviewed October 15, 2026. The examples are original exercises, not results from paid accounts or a completed memory benchmark.
First decide whether the fact changed
Consider Mara, an original adult bookshop character. At the start of a story, the shop opens on Friday. Later, you change the plan to Saturday. Two different corrections are possible:
- The latest response misunderstood a Saturday plan that was already clear.
- An older, still-active setup field continues to say Friday.
In the first case, repair the response or clarify the immediate scene. In the second, look for the stored instruction. Repeatedly discussing the contradiction can leave more competing statements in the transcript without making the authoritative version clearer.
Keep the replacement unambiguous: “The bookshop opens on Saturday. Friday was the earlier plan.” That preserves the meaning of an old conversation without pretending Friday was never mentioned.
Nomi separates ongoing context from the current situation
Nomi’s Backstory+ documentation names sections for core context, expression, current roleplay, preferences and boundaries, among others. Those labels help you decide whether the changed detail is part of a continuing setup or only the current scene.
For Mara, Current Roleplay is the documented section for the shop’s current opening plan. A recurring nickname goes in Nicknames, while how the character expresses herself belongs in Inclination. Put “The bookshop opens on Saturday; Friday was the earlier plan” in the current-situation note rather than treating a temporary schedule as a permanent personality trait. These are named sections from the guide, not invented form fields.
Nomi’s memory explanation also distinguishes memory from successful retrieval. A missed answer is not sufficient evidence that a detail has been deleted. Its suggested approach is to provide relevant context or a more specific reminder. That is different from interrogating the character with an unexplained “Do you remember?”
Kindroid distinguishes an evolving summary from a durable note
Kindroid’s memory guide documents subscriber-visible Learned Context and separates it from persistent setup and retrievable memories. Learned Context can be edited, but its summaries can change as conversation develops. For a fact that should remain important, the guide directs readers toward Key Memories or an appropriate journal entry.
That creates a useful editing question: are you correcting what the service has summarized, or defining something you deliberately want to preserve? Those actions can overlap, but treating them as identical hides an important distinction.
The same guide describes relevance-based retrieval and keyword-triggered journal behavior. A detail being stored does not mean it must appear in every reply. Do not infer perfect compliance from the existence of a memory control.
For this changed plan, inspect Personality → Learned Context → Ongoing threads for a stale summary. If you choose a journal instead, use a specific keyphrase such as “Mara’s bookshop opening” and the corrected date in the entry. The guide says keyphrase matching uses the user’s messages, not the AI’s response. A subsequent question must therefore include that phrase if you are deliberately checking that journal route.
A small, interpretable check
Use this proposed exercise after making one correction:
- Record the old statement, the replacement and the field you edited.
- Ask a natural question that needs the replacement: “For Mara’s bookshop opening, which tasks should we finish the day before?”
- Continue briefly on another topic, then return with a relevant cue.
- Note whether the reply used the new date, repeated the old one or avoided the date entirely.
Avoid changing the personality, model settings and opening scene during the same check. If several things change together, a better reply cannot tell you which change helped. A successful answer with the corrected date still visible in recent chat establishes less than a later answer without that immediate cue.
Choose Nomi’s documented section approach if you want a clearly separated relationship and scene brief. Inspect Kindroid’s memory controls if you want to distinguish an evolving summary from explicitly maintained context. For either service, make one specific correction and retain a note of what you changed; do not mistake a polished replacement sentence for proof that every underlying memory has changed.